How microsoft/moge works
MoGe estimates monocular 3D geometry from images, returning point, depth, and normal maps and camera FOV. The main user workflow is to load a pretrained model and run inference; CLI, panorama, and interactive-app interfaces provide additional ways to use it. The repository also includes data preparation, losses, and training/checkpoint workflows. Edges are limited to relationships stated in the README or shown by the supplied source excerpts and SOURCE INDEX; some interface and internal details are unsampled.
Main entry points: __init__.py Model registry · v2.py MoGe v2 · train_moge3.py Training runners · app.py Interactive app
Components 22
User interfaces
- Inference CLI
moge/scripts/cli.py - Image inference
moge/scripts/infer.py - Panorama inference
moge/scripts/infer_panorama.py - Interactive app
moge/scripts/app.py - 3D viewer
moge/utils/gradio_3d_viewer/gradio_3d_viewer.py
Geometry models
- Model registry
moge/model/__init__.py - MoGe v1
moge/model/v1.py - MoGe v2
moge/model/v2.py - MoGe v3
moge/model/v3.py - DINOv2 encoder
moge/model/modules/dinov2_encoder.py - Convolution neck
moge/model/modules/conv_stack.py - Sparse refinement
moge/model/modules/sparse_unet.py
Geometry utilities
- Image and map I/O
moge/utils/io.py - Tensor geometry
moge/utils/geometry_torch.py - Array geometry
moge/utils/geometry_numpy.py
Training system
- Training runners
moge/train/train_moge3.py - Training data
moge/train/dataloader.py - Training losses
moge/train/losses.py - Checkpoint manager
moge/train/checkpoint.py - Training visualization
moge/train/visualization.py
Other components
- Image user
- Hugging Face Hub
Connections 25
- Image user → Inference CLI: runs inference
README.md - Inference CLI → Image inference: dispatches to
- Image inference → Model registry: loads model
- Image user → Interactive app: uses demo
- Interactive app → Model registry: loads models
moge/scripts/app.py - Interactive app → Image and map I/O: reads and writes
moge/scripts/app.py - Interactive app → 3D viewer: uses viewer
moge/scripts/app.py - Panorama inference → Model registry: loads models
- Model registry → MoGe v1: exports version
moge/model/__init__.py - Model registry → MoGe v2: exports version
moge/model/__init__.py - Model registry → MoGe v3: exports version
moge/model/__init__.py - Hugging Face Hub → Model registry: provides weights
README.md - MoGe v1 → Tensor geometry: uses geometry
moge/model/v1.py - MoGe v2 → Tensor geometry: uses geometry
moge/model/v2.py - MoGe v2 → DINOv2 encoder: constructs encoder
moge/model/v2.py - MoGe v2 → Convolution neck: constructs neck
moge/model/v2.py - MoGe v3 → MoGe v2: extends model
moge/model/v3.py - MoGe v3 → Sparse refinement: uses refinement
moge/model/v3.py - Training runners → Training data: builds data pipeline
moge/train/train_moge3.py - Training runners → Training losses: computes losses
moge/train/train_moge3.py - Training runners → Checkpoint manager: manages checkpoints
moge/train/train_moge3.py - Training runners → Training visualization: renders training results
moge/train/train_moge3.py - Training losses → Tensor geometry: uses geometry
moge/train/losses.py - Training data → Array geometry: uses geometry
moge/train/dataloader.py - Training data → Image and map I/O: reads data
moge/train/dataloader.py